org.apache.hcatalog.data.schema.HCatSchema.remove(HCatFieldSchema hcatFieldSchema) makes the following call:
but fieldPositionMap is of type Map<String, Integer> so the element is not getting removed
Here's a detailed comment from Sushanth Sowmyan
The result is that that the name will not be removed from fieldPositionMap. This results in 2 things:
a) If anyone tries to append a field to a hcatschema after having removed that field, it shouldn't fail, but it will.
b) If anyone asks for the position of the removed field by name, it will still give the position.
Now, there is only one place in hcat code where we remove a field, and that is called from HCatOutputFormat.setSchema, where we try to detect if the user specified partition column names in the schema when they shouldn't have, and if they did, we remove it. Normally, people do not specify this, and this check tends to be superfluous.
Once we do this, we wind up serializing that new object (after performing some validations), and this does appear to stay through the serialization (and eventual deserialization) which is very worrying.
However, we are luckily saved by the fact that we do not append that field to it at any time(all appends in hcat code are done on newly initialized HCatSchema objects which have had no removes done on them), and we don't ask for the position of something we do not expect to be there(harder to verify for certain, but seems to be the case on inspection).
The main part that gives me worry is that HCatSchema is part of our public interface for HCat, in that M/R programs that use HCat can use it, and thus, they might have more interesting usage patterns that are hitting this bug.
I can't think of any currently open bugs that is caused by this because of the rarity of the situation, but nevertheless, something we should fix immediately.